FL-SEHGT: Federated Learning-Based Heterogeneous Graph Transformer with Squeeze-and-Excitation for ASD Identification

Autism spectrum disorder (ASD) is a neurodevelopmental disorder associated with widespread functional brain changes. Although deep learning has advanced computer-aided diagnosis of ASD using neuroimaging data, three key challenges remain: limited classification accuracy due to inadequate feature extraction, underexplored model interpretability that impedes understanding of the underlying neural mechanisms, and privacy concerns over multi-site data sharing that constrain practical deployment of collaborative diagnostic models. To address these challenges, we propose a federated learning-based squeeze-and-excitation heterogeneous graph transformer (FL-SEHGT) that utilizes resting-state functional MRI (rsfMRI) data to improve diagnostic accuracy relative to the evaluated federated baselines, enhance interpretability, and enable collaborative multi-site training without centralizing raw neuroimaging data. Experimental results on 871 subjects from 17 ABIDE I sites show that FL-SEHGT achieves a mean site accuracy of 63.45% across all 17 sites and 70.54% across the 10 sites with more than 40 subjects each, outperforming graph-augmentation-guided federated knowledge distillation (GAFD) (60.11%, 58.63%) and local-global federated learning (LG-FedAvg) (52.61%, 57.11%) on the corresponding site groups. The 10-site result also surpasses traditional models trained independently at each site (which achieved accuracies of 63.4%, 55.7%, and 60.2%). In a pooled five-fold evaluation, the proposed SEHGT attains an AUC of 71.7%. Altered functional connectivity patterns predominantly involving the default mode, cerebellar, and executive-control networks are identified as candidate neuroimaging features associated with ASD diagnosis. In summary, FL-SEHGT offers an effective, interpretable, and privacy-preserving framework for neuroimaging-based computer-aided diagnosis of ASD.